Le 28/09/2011 12:04, Justin Albstmeijer a écrit : > Hi aoe users and developers. > > I'm testing a setup that consists of two storage servers, which > replicate lvm volumes using drbd in master-master mode. > > Both storage servers export the same drbd/aoe block devices using qaoed > or ggaoed. > > The aoe kernel module, on the client machines witch mount the aoe block > devices (aoe devices are exclusively mounted on a server), load balances > between the two storage servers. > > All seems to work fine and performance is acceptable. Wow, I sounds really tricky to me... not sure I would do this in an active-active mode like you did. Can anybody confirm exporting the same block device (well... replicated block devices) from 2 diffrenets servers is fair/clever/secure? Does the aoe kernel module really deals with load balancing over multiple targets? How do you export your block device? same lvs have same shelf/slot numbers on both servers?
> The only worry I have is that qaoed or ggaoed might buffer the writes > before committing them to drbd, causing inconstancy in the replication. Indeed > This could be a problem in normal operation, but surely if one of the > storage servers would power-off unexpectedly without committing all it's > writes to drbd. > > Am I right to worry about this?. I am, just reading your post. :) > Should for this reason direct-io be enabled in the qaoed or ggaoed > configuration?. I have not tested the performance impact yet on this > setup, but from other aoe tests I would expect a sharp decrease in > performance. direct-io is mandatory in your case, as well as a raid controller with BBU... at least. Of course it has serious impact on performances, but combined with synchronous drdb replication (which I guess is the chosen replication method here), I don't know if it's sensitive. > Should I consider not exporting the same drbd/aoe on each storage server > or investigate if the aoe kernel module can work in fail-over mode to > limit the possible impact of this non-committed/lost data still in the > buffers? I'd go the safer way: active/passive redundancy. > Any advise/feedback is welcome. > > Justin > > > > > ------------------------------------------------------------------------------ > All the data continuously generated in your IT infrastructure contains a > definitive record of customers, application performance, security > threats, fraudulent activity and more. Splunk takes this data and makes > sense of it. Business sense. IT sense. Common sense. > http://p.sf.net/sfu/splunk-d2dcopy1 > _______________________________________________ > Aoetools-discuss mailing list > Aoetools-discuss@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/aoetools-discuss -- <http://www.horoa.net> Alexandre Chapellon Ingénierie des systèmes open sources et réseaux. Follow me on twitter: @alxgomz <http://www.twitter.com/alxgomz> ------------------------------------------------------------------------------ All the data continuously generated in your IT infrastructure contains a definitive record of customers, application performance, security threats, fraudulent activity and more. Splunk takes this data and makes sense of it. Business sense. IT sense. Common sense. http://p.sf.net/sfu/splunk-d2dcopy1 _______________________________________________ Aoetools-discuss mailing list Aoetools-discuss@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/aoetools-discuss
